The document describes a study that used a self-attention neural network architecture for human activity recognition using smartphone sensor data. The study compared the proposed self-attention model to convolutional neural network (CNN) and long short-term memory (LSTM) baselines. The self-attention model achieved a test accuracy of 91.75% for classifying six human activities, which was comparable to the baseline models. The study investigated components of the self-attention model like dropout rate, positional encoding, and scaling factors to determine the best performing model.

1. International Journal of Electrical and Computer Engineering (IJECE)
Vol. 13, No. 2, April 2023, pp. 2023~2029
ISSN: 2088-8708, DOI: 10.11591/ijece.v13i2.pp2023-2029  2023
Journal homepage: http://ijece.iaescore.com
Human activity recognition with self-attention
Yi-Fei Tan, Soon-Chang Poh, Chee-Pun Ooi, Wooi-Haw Tan
Faculty of Engineering, Multimedia University, Cyberjaya, Malaysia
Article Info ABSTRACT
Article history:
Received Mar 31, 2022
Revised Sep 15, 2022
Accepted Oct 13, 2022
In this paper, a self-attention based neural network architecture to address
human activity recognition is proposed. The dataset used was collected using
smartphone. The contribution of this paper is using a multi-layer multi-head
self-attention neural network architecture for human activity recognition and
compared to two strong baseline architectures, which are convolutional
neural network (CNN) and long-short term network (LSTM). The dropout
rate, positional encoding and scaling factor are also been investigated to find
the best model. The results show that proposed model achieves a test
accuracy of 91.75%, which is a comparable result when compared to both
the baseline models.
Keywords:
Convolution neural network
Human activity recognition
Long short-term memory
Self-attention
This is an open access article under the CC BY-SA license.
Corresponding Author:
Yi-Fei Tan
Faculty of Engineering, Multimedia University
63100 Cyberjaya, Selangor, Malaysia
Email: yftan@mmu.edu.my
1. INTRODUCTION
Around the world, the population of elderly is increasing rapidly. According to United Nations
analysis of recent trends of elderly population, the population is predicted to increase to 2.1 billion by 2050
[1]. Department of Statistics Malaysia projected the elderly population to reach about 20% of total population
[2]. Therefore, there are research which focused on improving elderly via assisted living technology [3]. One
of the key technologies is activity recognition. Activity recognition is a task of recognizing human activities
based on a series of observations of human actions [4]. Computer scientists addressed this task by using
various machine learning algorithms which feed on the observational data of human actions and classify the
given data into specific activity type. The observational data of activities can be collected by using different
means such as sensors or cameras [5]–[7]. In this paper, the focus is on sensor-based activity recognition
algorithm because sensory-based system is much less intrusive and lightweight when compared to
vision-based methods. On-body sensors such as accelerometer and gyroscope can be used to collect
acceleration and angular velocity of human body in various axis [8]–[11]. On-body sensors as data collectors
are ubiquitous and prevalent in human activity recognition research because of their relatively low cost. For
example, electronic devices such as smartphones, smart watches and wearable activity tracker have
embedded accelerometer and gyroscope. Besides, they are not limited to any location because people can
carry them everywhere. One of the limitations of using on-body sensors is the battery issue. The pattern of
the time series sensory data for each type of activity is distinct. For example, time series acceleration of
running should have a higher rate of change compared to walking. This unique distinction in pattern allows
machine learning algorithm to classify them.
The epplication of neural network architecture in human activity recognition has grown in recent
years. Besides neural networks, there are several machine learning algorithms that were used for human
activity recognition which include support vector machine (SVM) [12] and random forest (RF) [13]. Anguita
et al. [12] used a dataset collected using smartphone’s inertial sensors. In this research, SVM which exploits

2.  ISSN: 2088-8708
Int J Elec & Comp Eng, Vol. 13, No. 2, April 2023: 2023-2029
2024
fixed-point arithmetic was used as a multiclass classifier. Contrary to traditional SVM, this approach
demonstrated a considerable improvement in computational cost while achieving similar accuracy. Xu et al.
[13] proposed a RF model which classify human activity based on sensory data from wearable devices. RF
model managed to obtain an overall accuracy of 90%.
The prevalent neural network architecture includes CNN and recurrent neural network (RNN).
Wu and Zhang [14] collected a 128-dimensional time series sensory data from accelerometer and gyroscope
of a smartphone. In this study, they trained an 8-layer CNN network to classify the sensory data into type of
activities. Their method achieved comparable results with conventional and state of the art models. Xi et al.
[15] reasoned using pooling after convolution to expand the receptive fields is not advisable because it could
bring about information loss. They proposed an alternative model based on dilated convolution which can
expand receptive field without compromising on information loss.
Long-short term network (LSTM) [16] is a widely used variant of RNN. Güney and Erdas [17] used
LSTM to classify raw sensory data collected from three-axis accelerometer. The proposed method managed
to obtain a classification accuracy of 91.34%. Mahmud et al. [18] used a multi-stage LSTM to extract
temporal features from sensory data and achieved a F1 score of 83.9%. Each stage of LSTM is used to
extract features from each sensor’s data. Then, the output of every stage of LSTM are aggregated using fully
connected layers.
There are also research works which combine both CNN and RNN for human activity recognition.
Mutegeki and Han [19] introduced a hybrid model of CNN and LSTM which achieved high classification
accuracy and reduced the complexity of the model. In [20], a hybrid model based on CNN and gated
recurrent unit (GRU) is proposed. GRU is another variant of RNN. In this work, the authors trained the
model on recognition of pairwise similar activities. This means that this model can be trained on data of one
type of activity and classify the data of another similar activity type.
In this work, we developed a human activity classification algorithm based on multi-head
self-attention neural network architecture. We trained and evaluated our models on a dataset which consists
of six activities. The proposed model was compared with two strong baselines based on CNN and LSTM
respectively. This work consists of ablation study of various component that is commonly used in
self-attention model such as dropout, scaling study and positional encoding. Our results show that the
proposed self-attention neural network architecture delivered a test accuracy on unseen data which is on-par
with both the baseline architectures. The organization of this paper is: in section 2 descriptions of the dataset
are given; in section 3 we present the detailed architecture of baseline CNN and LSTM model, as well as
proposed self-attention model; section 4, the experimental result and discussion are discussed; section 5, we
conclude the paper with a conclusion and future work.
2. DATASET
The dataset used in this research is a publicly available human activity recognition dataset called
UCI-HAR dataset [21]. This data is collected from 30 person aged 19 to 48 years old. Each volunteer was
asked to carry out six types of activities (walking, walking upstairs, walking downstairs, standing, sitting, and
laying) wearing a Samsung Galaxy smartphone on the waist. The dataset is not raw sensory data of tri-axis
accelerometer and gyroscope. It was pre-processed with noise filters and sampled with a sliding window of
2.56 seconds with 50% overlap. Eventually, each instance of sensory data consists of 128 readings or time
steps. The dataset consists of 10,299 instances in total. The authors of the dataset split the dataset into 70%
training set and 30% test set based on the volunteers. This means that volunteers for training set are not
included in test set and vice versa. The percentage of how much each activity type made up the train and test
set and other details of the dataset is summarized in Table 1.
Table 1. Class percentage for train and test set
Activity Type Train (%) Train (%)
Walking 16.68 16.83
Walking upstairs 14.59 15.98
Walking downstairs 13.41 14.25
Sitting 17.49 16.66
Standing 18.69 18.05
Laying 19.14 18.22
Total number of instances 7352 2947
There are nine features available for each instance of data which include triaxial acceleration from
the accelerometer (total acceleration), the estimated body acceleration and the triaxial angular velocity from

3. Int J Elec & Comp Eng ISSN: 2088-8708 
Human activity recognition with self-attention (Yi-Fei Tan)
2025
the gyroscope. Each of these three categories contribute 3 features in x, y and z axis. Each instance of data
consists of 128 time steps and 9 dimensions per time step for each of the features and is a matrix of size
128×9.
3. RESEARCH METHOD
Time series classification problem is usually solved using certain type of RNN architecture such as
LSTM and GRU. However, RNN computation cannot be done in parallel for all the time steps. RNN
architecture design requires computation to be computed sequentially. Due to this fact, CNN is explored as
an alternative for time series data because CNN can compute all the time steps in parallel with convolution.
In this work, we developed two baselines, one based on LSTM and another based on CNN. We compared our
multi-head self-attention model to these baselines. Self-attention models are widely used model in natural
language processing (NLP) tasks which are also sequential in nature.
3.1. CNN architecture
The CNN baseline model has 2 layers of 1D convolutional layer with filter of size 3. The number of
filters for both layers are 64. Both of the convolutional layers use non-linear activation function called
rectified linear unit (ReLU). The output of second convolutional layer is passed to a 1D max pooling layer
with pooling size of 2. The output is then flattened and connected to two layers of fully connected layers. The
first fully-connected layer has 128 hidden units and uses ReLU activation function. The final layer has 6 units
corresponding to 6 classes of activities and uses a SoftMax activation function. Figure 1 visualized the CNN
model which serves as one of our baseline models.
3.2. LSTM architecture
The LSTM baseline model has 3 layers of LSTM. The first LSTM layer has 32 hidden units. The
second LSTM layer has 64 hidden units. The third LSTM layer has 32 hidden units. The final time step of the
third LSTM layer is connected to a fully connected layer. The fully connected layer has 6 units corresponding
to 6 classes of activities and uses a SoftMax activation function. Figure 2 summarizes the LSTM model we
used.
Figure 1. CNN model Figure 2. LSTM model
3.3. Multi-head self-attention architecture
3.3.1. Multi-head self-attention layer
Recently, self-attention neural network architecture [22] is the go-to neural network architecture for
many NLP tasks. Large language models with lots of self-attention layers such as bidirectional encoder
representations from transformers (BERT) can be trained on large text corpus without annotations [23]. Then,
it acts a pre-trained model for downstream NLP tasks such as question answering (QA) [24]. Recently,
researchers are investigating the possibility of using attention-based model for computer vision tasks such as
image classification [25] and object detection [26].
In [22], Google scientists introduced fully attention-based sequence-to-sequence model called
transformer. Transformer consists of a basic building block of self-attention called scaled dot product
attention which is given by equations listed at (1). The input sequence, X is a tensor of N×T×F where N is the
number of data examples, T is the number of time steps and F is the number of features. For UCI-HAR
dataset, T=128 and F=9. The input X is passed through three fully connected layer to produce Q, K ad V. All
three fully connected layers are linear without any non-linear activation function. Their parameters include
the weights denoted as Wq, Wk, Wv and biases denoted as bq, bk and bv. Each of the weight term is a matrix of
F×H where H is the number of hidden units of the fully connected layer. In addition, each of the bias term is
a vector of H dimension. These fully connected layers produce three output tensors denoted by Q, K and V.

4.  ISSN: 2088-8708
Int J Elec & Comp Eng, Vol. 13, No. 2, April 2023: 2023-2029
2026
Each of these output tensors is a tensor of N×T×H. We then carry out dot product for both Q and K with a
scaling factor of
1
√dk
.
Q = XWq + bq
K = XWk + bk
V = XWv + bv
Attention(Q,K,V)=SoftMax (
Q ∙KT
√dk
)V (1)
The output is then passed through a SoftMax activation and multiplied with V which produce a final tensor of
N×T×H.
3.3.2. Multi-head self-attention model
Figure 3 shows the multi-head self-attention model for human activity recognition for one data
instance. A data instance of 128×9 is expanded in dimension to 128×128 after passing through a fully
connected layer (labelled FC in Figure 3) with ReLU activation. Then, the matrix of 128×128 is passed to
2 blocks of “4-head self-attention block”. Each 4-head self-attention block consists of 2 layers. The first layer
is a 4-head self-attention layer described in Part 1 of this subsection. Each individual self-attention layer
outputs a 128×128 matrix. The 4-head self-attention layer results in four 128×128 matrices which is then
concatenated to a matrix of 128×512. The output matrix of this concatenated matrix is passed to a fully
connected layer with 128 hidden units, which then outputs a 128×128 matrix. The output matrix of 2 blocks
of self-attention is matrix of 128×128, which is then unrolled to a column vector of 16384 dimensions. The
column vector is then passed through a FC layer with 128 units using ReLU and then to the prediction layer
with six units with SoftMax activation. The output is a vector of 6. Finally, we added dropout for some FC
layer. Table 2 summarizes the model and listed the number of parameters for each layer.
Figure 3. Multi-head self-attention model used for time series activity recognition
Table 2. Model summary
Layer Output Shape Number of parameters
FC (128 units, ReLU) 128×128 1280
4-head attention 4×128×128 4×49,536
Concatenate 128×512 -
FC (128 units, ReLU, dropout) 128×128 65,664
4-head attention 4×128×128 4×49,536
Concatenate 128×512 -
FC (128 units, ReLU, dropout) 128×128 65,664
Flatten 16384 -
FC (128 units, ReLU, dropout) 128 2,097,280
FC (6 units, Softmax) 6 774
FC (128 units, ReLU) 128×128 1,280
Total - 2,626,950

5. Int J Elec & Comp Eng ISSN: 2088-8708 
Human activity recognition with self-attention (Yi-Fei Tan)
2027
3.3.3. Positional encoding
Self-attention layer does not have any convolution and recurrence. For an instance of data, LSTM
takes in each time step of the data recurrently. As for CNN, each position of the output feature map of
convolutional layer corresponds to a receptive field in the input. Both LSTM and CNN have location
information of each time step of the sequential input. On the contrary, the output at each time step of the
self-attention layer is independent of computation at other time steps. This results in self-attention is not
location or positionally aware. For example, if the input is the word “apple”. To self-attention, “apple” or its
anagrams such as “papel” are the same. Therefore, the authors proposed that it is vital to inject some
positional information to the input.
In this work, we tried the positional encoding function introduced in [22]. The positional encoding
function generates a matrix PE of same shape as input X. The positional information is injected using
elementwise addition of X and PE as given in (2).
𝑋 = 𝑋 + 𝑃𝐸 (2)
4. RESULTS AND DISCUSSION
We evaluated the proposed self-attention on test set using the evaluation metric, accuracy. For every
experiment, we used Adam optimizer and run the training for 50 epochs. We picked the best test accuracy out
of all the 50 epochs to be presented.
4.1. Dropout
Dropout is a regularization method [27]. We added dropout at some of the FC layer. The detailed
position of dropout in the self-attention model is given in Table 2. Since the dropout rate is a hyperparameter,
we have tried out a few choices of dropout rate. Table 3 listed the test accuracy of our self-attention model
for difference choices of dropout rate. It was discovered that self-attention model with dropout rate of 0.2 has
the highest test accuracy of 0.9175. Therefore, we fixed dropout rate to 0.2 for later experiments.
Table 3. Choice of dropout rate and test accuracy
Dropout Rate Test Accuracy
None 0.9070
0.1 0.9152
0.2 0.9175
0.3 0.9104
0.5 0.9023
4.2. Positional encoding
Experiment was carried out to investigate positional encoding. Our experiment result is tabulated in
Table 4. Addition of positional encoding results in a test accuracy of 0.8985 which is less accurate than the
model without addition of positional encoding. Positional encoding does not improve the performance of our
proposed self-attention model. We believed the reason behind this is due to the sequence length of 128 for
this dataset is relatively shorter compared to NLP tasks. Besides, we conjecture that the activity recognition
dataset is sensitive to noise. Addition of positional encoding may have brought about a regularizing side
effect.
4.3. Scaling factor
In this part, we evaluated the necessity of scaling factor. Scaling factor
1
√dk
is part of the
self-attention layer given in (1). For our proposed model, dk=128. By setting dk=1, the scaling factor’s effect
disappear as it is multiplied by 1. We discovered adding scaling factor will hurt the accuracy of the model.
Table 5 shows the effect of scaling factor on test accuracy. Addition of scaling factor results in a model with
test accuracy of 0.8853 which is lower than a model without scaling factor. Therefore, the model that was
used to compare with the two baselines did not include scaling factor.
Table 4. Effects of positional encoding test accuracy
With Positional Encoding Test Accuracy
Yes 0.8985
No 0.9175
Table 5. Effects of scaling factor on test accuracy
dk Train Accuracy Test Accuracy
1 0.9643 0.9175
128 0.9519 0.8853

6.  ISSN: 2088-8708
Int J Elec & Comp Eng, Vol. 13, No. 2, April 2023: 2023-2029
2028
4.4. Comparison with baseline models
In this paper, we introduced two baseline models which include CNN and LSTM. We compared the
performance of our proposed self-attention model with these two baselines. Table 6 listed the train and test
accuracies of self-attention model and two baseline models: CNN and LSTM. The model with the highest
test accuracy is LSTM followed by CNN. Our proposed model has the lowest test accuracy of 0.9175 with
less than 1% difference from LSTM’s test accuracy of 0.9237. Our proposed model is comparable to both the
baseline models.
Table 6. Train and test accuracy for self-attention model, CNN and LSTM
Model Train Accuracy Test Accuracy
CNN 0.9536 0.9223
LSTM 0.9559 0.9237
Self-attention 0.9643 0.9175
5. CONCLUSION AND FUTURE WORK
In this paper, an activity recognition algorithm using multi-head self-attention neural network
architecture is introduced. The experiment to evaluate this proposed model was conducted using UCI-HAR
dataset. The results demonstrated that the proposed model has a test accuracy of 0.9175 which is comparable
to two baseline models: CNN and LSTM. In addition, we also discovered some of the common components
of self-attention layer such as scaling factor and positional encoding actually reduce test accuracy. In future
work, we would like to apply our model on other datasets and collect our own dataset.
REFERENCES
[1] WHO, “Ageing and health,” World Health Organization, 2021. https://www.who.int/news-room/fact-sheets/detail/ageing-and-
health (accessed Oct. 04, 2021).
[2] M. U. Mahidin, Selected demographic indicators. Department of Statistics Malaysia, 2018.
[3] G. Cicirelli, R. Marani, A. Petitti, A. Milella, and T. D’Orazio, “Ambient assisted living: a review of technologies, methodologies
and future perspectives for healthy aging of population,” Sensors, vol. 21, no. 10, May 2021, doi: 10.3390/s21103549.
[4] A. K. S. Kushwaha, O. Prakash, A. Khare, and M. H. Kolekar, “Rule based human activity recognition for surveillance system,”
in 2012 4th International Conference on Intelligent Human Computer Interaction (IHCI), Dec. 2012, pp. 1–6, doi:
10.1109/IHCI.2012.6481853.
[5] M. C. Sorkun, A. E. Danisman, and O. D. Incel, “Human activity recognition with mobile phone sensors: Impact of sensors and
window size,” in 2018 26th Signal Processing and Communications Applications Conference (SIU), May 2018, pp. 1–4, doi:
10.1109/SIU.2018.8404569.
[6] A. Bagate and M. Shah, “Human activity recognition using RGB-D sensors,” in 2019 International Conference on Intelligent
Computing and Control Systems (ICCS), May 2019, pp. 902–905, doi: 10.1109/ICCS45141.2019.9065460.
[7] S. Tao, M. Kudo, H. Nonaka, and J. Toyama, “Camera view usage of binary infrared sensors for activity recognition,” in
Proceedings of the 21st International Conference on Pattern Recognition (ICPR2012), 2012, pp. 1759–1762.
[8] T. Kaytaran and L. Bayindir, “Activity recognition with wrist found in photon development board and accelerometer,” in 2018
26th Signal Processing and Communications Applications Conference (SIU), 2018, pp. 1–4, doi: 10.1109/SIU.2018.8404630.
[9] S-M. Lee, S. M. Yoon, and H. Cho, “Human activity recognition from accelerometer data using convolutional neural network,” in
2017 IEEE International Conference on Big Data and Smart Computing (BigComp), Feb. 2017, pp. 131–134, doi:
10.1109/BIGCOMP.2017.7881728.
[10] P. Gupta and T. Dallas, “Feature selection and activity recognition system using a single triaxial accelerometer,” IEEE
Transactions on Biomedical Engineering, vol. 61, no. 6, pp. 1780–1786, Jun. 2014, doi: 10.1109/TBME.2014.2307069.
[11] N. Hardiyanti, A. Lawi, Diaraya, and F. Aziz, “Classification of human activity based on sensor accelerometer and gyroscope
using ensemble SVM method,” in 2018 2nd East Indonesia Conference on Computer and Information Technology (EIConCIT),
Nov. 2018, pp. 304–307, doi: 10.1109/EIConCIT.2018.8878627.
[12] D. Anguita, A. Ghio, L. Oneto, X. Parra, and J. L. Reyes-Ortiz, “Human activity recognition on smartphones using a multiclass
hardware-friendly support vector machine,” in Lecture Notes in Computer Science, Springer Berlin Heidelberg, 2012,
pp. 216–223.
[13] L. Xu, W. Yang, Y. Cao, and Q. Li, “Human activity recognition based on random forests,” in 2017 13th International
Conference on Natural Computation, Fuzzy Systems and Knowledge Discovery (ICNC-FSKD), Jul. 2017, pp. 548–553, doi:
10.1109/FSKD.2017.8393329.
[14] W. Wu and Y. Zhang, “Activity recognition from mobile phone using deep CNN,” in 2019 Chinese Control Conference (CCC),
Jul. 2019, pp. 7786–7790, doi: 10.23919/ChiCC.2019.8865142.
[15] R. Xi, M. Hou, M. Fu, H. Qu, and D. Liu, “Deep dilated convolution on multimodality time series for human activity
recognition,” in 2018 International Joint Conference on Neural Networks (IJCNN), Jul. 2018, pp. 1–8, doi:
10.1109/IJCNN.2018.8489540.
[16] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural Computation, vol. 9, no. 8, pp. 1735–1780, Nov. 1997,
doi: 10.1162/neco.1997.9.8.1735.
[17] S. Guney and C. B. Erdas, “A deep LSTM approach for activity recognition,” in 2019 42nd International Conference on
Telecommunications and Signal Processing (TSP), Jul. 2019, pp. 294–297, doi: 10.1109/TSP.2019.8768815.
[18] T. Mahmud, S. S. Akash, S. A. Fattah, W.-P. Zhu, and M. O. Ahmad, “Human activity recognition from multi-modal wearable
sensor data using deep multi-stage LSTM architecture based on temporal feature aggregation,” in 2020 IEEE 63rd International
Midwest Symposium on Circuits and Systems (MWSCAS), Aug. 2020, pp. 249–252, doi: 10.1109/MWSCAS48704.2020.9184666.

7. Int J Elec & Comp Eng ISSN: 2088-8708 
Human activity recognition with self-attention (Yi-Fei Tan)
2029
[19] R. Mutegeki and D. S. Han, “A CNN-LSTM approach to human activity recognition,” in 2020 International Conference on
Artificial Intelligence in Information and Communication (ICAIIC), Feb. 2020, pp. 362–366, doi:
10.1109/ICAIIC48513.2020.9065078.
[20] M. S. Siraj and M. A. R. Ahad, “A hybrid deep learning framework using CNN and GRU-based RNN for recognition of pairwise
similar activities,” in 2020 Joint 9th International Conference on Informatics, Electronics and Vision (ICIEV) and 2020 4th
International Conference on Imaging, Vision and Pattern Recognition (icIVPR), Aug. 2020, pp. 1–7, doi:
10.1109/ICIEVicIVPR48672.2020.9306630.
[21] D. Anguita, A. Ghio, L. Oneto, X. Parra, and J. L. Reyes-Ortiz, “A public domain dataset for human activity recognition using
smartphones,” in Proceedings of the 21th international European symposium on artificial neural networks, computational
intelligence and machine learning, 2013, pp. 437–442.
[22] A. Vaswani et al., “Attention is all you need,” Advances in neural information processing systems, Jun. 2017.
[23] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: pre-training of deep bidirectional transformers for language
understanding,” arXiv:1810.04805, Oct. 2018.
[24] D. Luo, J. Su, and S. Yu, “A BERT-based approach with relation-aware attention for knowledge base question answering,” in
2020 International Joint Conference on Neural Networks (IJCNN), Jul. 2020, pp. 1–8, doi: 10.1109/IJCNN48605.2020.9207186.
[25] A. Dosovitskiy et al., “An image is worth 16x16 Words: transformers for image recognition at scale,” arXiv:2010.11929, Oct.
2020.
[26] N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko, “End-to-end object detection with transformers,” in
Computer Vision ECCV 2020, Springer International Publishing, 2020, pp. 213–229.
[27] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, “Dropout: a simple way to prevent neural networks
from overfitting,” The journal of machine learning research, vol. 15, no. 1, pp. 1929–1958, 2014.
BIOGRAPHIES OF AUTHORS
Yi-Fei Tan received her B.Sc. (Hons), M.Sc. and Ph.D. from University of Malaya
(UM), Malaysia. She is currently a senior lecturer at the Faculty of Engineering in Multimedia
University (MMU), Cyberjaya, Malaysia. Her research areas include machine learning, deep
learning, image processing, big data analytics and queueing theory. She can be contacted at
email: yftan@mmu.edu.my.
Soon-Chang Poh obtained his B. Eng. (Hons.) Electronics and Master of
Engineering Science from Multimedia University, Malaysia. He is currently a research officer at
the Faculty of Engineering in Multimedia University (MMU), Cyberjaya, Malaysia. His research
interests include deep learning and anomaly detection. He can be contacted at email:
psoonchang@gmail.com.
Chee-Pun Ooi obtained his Ph.D. in Electrical Engineering from University of
Malaya, Malaysia, in year 2010. His current position is an Assoc. professor the Faculty of
Engineering in Multimedia University, Malaysia. He is a chartered engineer registered with
Engineering Council (the British regulatory body for Engineers), member of IET. His research
areas are FPGA based embedded systems and embedded systems. He can be contacted at email:
cpooi@mmu.edu.my.
Wooi-Haw Tan received his M.Sc. in Electronics from Queen’s University of
Belfast, UK and a Ph.D. in Engineering from Multimedia University. He is currently a senior
lecturer at Multimedia University. Dr. Tan’s areas of expertise include image processing,
embedded system design, internet of things (IoT), machine learning and deep learning. He can be
contacted at email: twhaw@mmu.edu.my.